The skripts now automatically switch between cda and conll format based
on the file ending.

TurboParser output (unused nodes are attached to a special "unused" node
instead of NIL) can be evaluated with the -u switch.

Logging is done with the logging framework.

Predictability computation has been factored out, support for english is
available.

argument parsing is done with argparse

Prefixes can be padded with virtual nodes to train TurboParser

Prefixes can be created with only padded nodes (useful for TP
input). Could be done with much less (CPU-)work but the infrastructure
was already in place.

folding and implicit are no longer supported since they don't map well
to other languages.